How to Run DeepSeek Locally (2026): Which Model Fits Your PC, Ollama, LM Studio and V4-Flash

“Run DeepSeek locally” sounds simple, but most guides skip the awkward truth: the newest DeepSeek V4 models are hundreds of gigabytes, and the small “DeepSeek” models in Ollama are actually distilled Qwen and Llama models.

In this guide I explain which DeepSeek you can really run on your hardware, from an 8 GB laptop GPU to a 128 GB RAM workstation, with exact Ollama, LM Studio and llama.cpp steps. All sizes and dates are checked as of November 2026.

Which DeepSeek Can You Run? (Hardware Table)

Start here. Find your GPU memory (or Mac unified memory) and pick the matching model. Sizes are the Ollama download sizes or the Unsloth GGUF sizes.

Not sure what you have? Check with my GPU VRAM guide, and see how much RAM you need for local AI.

What “deepseek-r1:8b” Really Is

When DeepSeek released R1 in 2025, it also trained smaller open models to copy R1’s step-by-step reasoning. These “distills” are what Ollama ships under the deepseek-r1 name.

Note: these distills are from 2025. They are still good at maths and logic puzzles, but newer local models often beat them. If you want the best model for a 16 GB card today, compare with Qwen3.8-27B on 16 GB and my best Ollama models list.

Method 1: Run DeepSeek With Ollama

Ollama is the fastest way to get DeepSeek running. It downloads the model, picks your GPU and gives you a chat in the terminal.

1. Install Ollama. Download it from ollama.com for Windows, Mac or Linux. On Linux run curl -fsSL https://ollama.com/install.sh | sh. Full steps are in my Ollama install guide.
2. Open a terminal. PowerShell or Command Prompt on Windows, Terminal on Mac and Linux.
3. Pull and run the model. Type ollama run deepseek-r1:8b. The 5.2 GB download starts and the chat opens when it finishes.
4. Ask a question. The model “thinks” first (you see its reasoning), then answers. Type /bye to exit.
5. Check it is on the GPU. In a second terminal run ollama ps. The PROCESSOR column should say 100% GPU.

Useful Ollama commands for DeepSeek

More commands are in my Ollama commands cheat sheet. To clear space later, see how to delete Ollama models.

Method 2: Run DeepSeek in LM Studio

LM Studio is a free desktop app with a point-and-click interface. It is a good choice if you would rather not use a terminal.

1. Install LM Studio from lmstudio.ai and open it.
2. Open the Discover tab (the magnifying glass) and search for DeepSeek R1.
3. Pick a size and quant. LM Studio marks files that fit your GPU. A Q4_K_M quant of the 8B or 14B model is the usual choice.
4. Download, then load it. Go to the Chat tab, choose the model at the top and start chatting.
5. Turn on GPU offload. In the model settings, set GPU offload to the maximum layers that fit.

Torn between the two apps? My Ollama vs LM Studio vs llama.cpp comparison helps you choose.

Add a ChatGPT-Style Interface

The Ollama terminal works, but a browser chat with history, file uploads and model switching is nicer. Open WebUI connects to Ollama in a few minutes.

• Open WebUI: full guide in my Open WebUI + Ollama setup.
• Ollama’s own app: the Windows and Mac installers include a simple chat window.
• Coding: plug local DeepSeek into VS Code; see DeepSeek in VS Code with Cline and Continue.

Method 3: Run the Real DeepSeek V4-Flash With llama.cpp

DeepSeek V4-Flash is the genuine V4 model: 284B total parameters with 13B active, 1M-token context and an MIT licence. It went on Hugging Face on 22 April 2026, and Unsloth publishes GGUF quants of the 0731 update that run in mainline llama.cpp.

V4-Flash GGUF sizes (Unsloth)

Note: V4 weights are already stored in FP4 and FP8, so they do not shrink as much as older models when quantized. That is why even 1-bit is over 80 GB.

Steps

1. Get llama.cpp. Download a recent release from github.com/ggml-org/llama.cpp, or build it: cmake -B build -DGGML_CUDA=ON then cmake --build build --config Release.
2. Install the Hugging Face CLI. Run pip install -U "huggingface_hub[cli]".
3. Download one quant. Run hf download unsloth/DeepSeek-V4-Flash-0731-GGUF --include "UD-IQ3_XXS/*" --local-dir models/v4-flash. Keep all shards in one folder.
4. Start the server with MoE offload. Point at the first shard and keep the expert layers in system RAM:

llama-server -m models/v4-flash/UD-IQ3_XXS/DeepSeek-V4-Flash-0731-UD-IQ3_XXS-00001-of-00004.gguf -ngl 99 --n-cpu-moe 40 -c 32768 --temp 1.0 --top-p 1.0 --min-p 0.0

Raise or lower the --n-cpu-moe number until the model fits your VRAM; a higher number keeps more experts in RAM. Open http://127.0.0.1:8080 for the built-in chat page.

What speed to expect

A desktop with a 16 to 32 GB NVIDIA GPU sits between those numbers, because the GPU handles attention and the busiest layers. The same offload approach is explained step by step in my Qwen3.8-Flash-Next on 2x 16 GB GPUs guide.

Tip: V4-Flash reasons by default. Add --reasoning off for quick answers, or --chat-template-kwargs '{"reasoning_effort":"max"}' for hard problems.

Can You Run V4.1-Flash or V4-Pro?

• DeepSeek V4.1-Flash-1017 (17 October 2026) has a 552B backbone plus 196B of “Engram” memory, with 8B to 16B active. Unsloth’s GGUFs run from 233 GB (1-bit) to 508 GB. The smallest needs at least 128 to 145 GB of RAM plus VRAM, with the Engram table streamed from a fast SSD.
• It also needs a llama.cpp build from a pull request (#30817) at the time of writing, not a normal release.
• DeepSeek V4-Pro is 1.6T parameters (49B active) and about 865 GB. It is a multi-GPU server model; DeepSeek’s reference code splits it across several processes.

For most people, the sensible way to use V4.1 or V4-Pro is the API, which is cheap. Prices are in my DeepSeek API pricing post, and code is in the DeepSeek API Python tutorial.

Ollama :cloud Tags Are Not Local

Ollama lists deepseek-v4.1-flash:cloud and deepseek-v4-pro:cloud. These run on Ollama’s servers, not your PC; your prompts are sent over the internet.

Ollama’s local deepseek-v4-flash library page says it was retired on 27 August 2026, so for a truly local V4 model, use llama.cpp as shown above.

Warning: if you chose local AI for privacy, do not use the :cloud tags. Check with ollama ps or the model name before pasting anything sensitive.

Settings and Speed Tips

• Temperature: R1 distills work best around 0.6. V4 models use 1.0 with top_p 1.0, as DeepSeek and Unsloth recommend.
• Context: Ollama uses a small default context. Raise it with /set parameter num_ctx 16384 inside the chat, but more context needs more VRAM.
• Keep it all on the GPU: if ollama ps shows a CPU/GPU split, pick a smaller tag. Speed drops sharply once layers spill into RAM.
• Two GPUs: Ollama and llama.cpp split models across both cards automatically.
• Thinking too long? Use --think=false in Ollama, or ask for “a short answer without showing your work”.

Privacy: Local vs the DeepSeek App

Troubleshooting

DeepSeek is very slow in Ollama

Run ollama ps. If it shows CPU use, the model is too big for your VRAM; switch to a smaller tag. If it shows 100% CPU on an NVIDIA PC, update your driver and see my Ollama not using GPU fix.

The model thinks forever or repeats itself

This is common with small reasoning models. Lower the temperature to about 0.6, raise num_ctx so the reasoning is not cut off, or use the 14B tag instead of 1.5B or 7B.

“model requires more system memory” in Ollama

Your RAM plus VRAM is too small for that tag. Close other apps, use a smaller tag, or add swap space on Linux.

llama.cpp crashes with CUDA out of memory on V4-Flash

Increase --n-cpu-moe, lower -c (context size), or pick a smaller quant. My CUDA out of memory guide has more fixes.

“unknown model architecture” in llama.cpp

Your llama.cpp is too old. Update to a current release for V4-Flash; V4.1-Flash still needs the PR build listed above.

The answer switches to Chinese

Distills sometimes mix languages. Add “Answer in English” to your prompt or a system prompt, and avoid very low temperatures.

My Take

For a normal PC, deepseek-r1:8b or 14b in Ollama is a fun, private reasoning model that takes five minutes to set up. Just know it is a distill, not the DeepSeek you use on the website.

If you have 128 GB of RAM and a decent GPU, V4-Flash in llama.cpp is the most exciting option: the real V4 architecture running at home. Everything bigger is honestly better through the API.

FAQ

Can I run DeepSeek offline?

Yes. After downloading a model with Ollama, LM Studio or llama.cpp, it runs with no internet connection. Only :cloud tags need the internet.

Is the DeepSeek in Ollama the same as the DeepSeek app?

No. Ollama’s deepseek-r1 tags up to 70b are distilled Qwen or Llama models. The app runs DeepSeek’s full V4-series models on its servers.

Which DeepSeek model is best for an 8 GB GPU?

deepseek-r1:8b (DeepSeek-R1-0528-Qwen3-8B, 5.2 GB). It leaves room for context on an 8 GB card.

Is running DeepSeek locally safe?

Yes. A local model is just a file of weights; it cannot send your data anywhere. Download from official sources such as Ollama, LM Studio or the Unsloth Hugging Face pages.

How much RAM do I need for DeepSeek V4?

V4-Flash needs about 83 GB (1-bit) to 162 GB (8-bit) of combined RAM and VRAM. V4.1-Flash needs at least 128 to 145 GB, and V4-Pro needs a server.

Can I use local DeepSeek for coding?

Yes. Connect Ollama to VS Code with Cline or Continue. For serious coding on 16 GB, newer coding models may do better than the 2025 distills.

Is DeepSeek free for commercial use?

DeepSeek’s own models use the MIT licence. Distills keep their base model’s licence: Apache 2.0 for Qwen bases and the Llama 3.3 licence for the 70b tag.

Sources

• Ollama library: deepseek-r1
• Ollama library: deepseek-v4.1-flash (cloud)
• Unsloth: DeepSeek-V4.1-Flash-1017 GGUF
• Unsloth docs: DeepSeek-V4.1-Flash
• ComputingForGeeks: Run DeepSeek V4 Flash locally
• Kevin Welter: Run DeepSeek locally? What V4 needs